An investigation of Frequentist and Ensemble Bayesian-aided techniques for prioritizing anomaly detection methods in time-series data
Vignesh Divakaran, Vipasha Rana · Decision Analytics Journal · 2025
Accurately detecting anomalous points in time-series data is critical, as false positives can mislead business stakeholders, waste valuable resources, and diminish the overall impact of the detection system. While various statistical and machine learning techniques are employed to flag potential anomalies, the challenge lies in evaluating the significance of each approach and refining the results to isolate definitive anomalies. This paper examines multiple anomaly tagging techniques and introduces novel weightage assignment methods to prioritize the most effective approaches, filtering out less reliable ones. Specifically, we explore two methods: simple Frequentist approach and Ensemble Bayesian-aided approach, with an emphasis on why the latter is particularly well-suited for anomaly detection. The proposed methodology is validated both theoretically and empirically on time-series datasets. Our findings demonstrate that the Ensemble Bayesian-aided approach significantly improves detection accuracy by accounting for future uncertainty and addressing edge case fallacies inherent in individual tagging methods. This research provides a robust framework for anomaly detection, offering a powerful solution that enhances precision and reliability across diverse applications. • Explore various anomaly tagging approaches and their effectiveness in data labeling. • Demonstrate the power of Bayesian methods for enhancing anomaly detection accuracy. • Compare frequentist and ensemble approaches in assigning statistical weightage. • Highlight the role of uncertainty in improving anomaly detection outcomes. • Showcases empirical validation of Bayesian-aided approaches on time-series datasets.